Voice Polling: The Sleeping Double Majority Curse

Yes trails about 45.0-55.0 in aggregated public polling (as of 13 Aug, pending new data)

No now leading on aggregate in every state, but Yes still ahead of national total in four states

Time for another Voice polling roundup.  There has been relatively little new data in the four weeks since the last update and this article includes some historic analysis that suggests that the current state polling picture for the Voice is actually highly unusual.  At the moment the state polling picture is irrelevant because No is ahead nationwide, but a benign state distribution is one thing Yes does have going for it should the national picture improve or if polls are underestimating Yes for some unprecedented reason.  What I find here is that it is almost unprecedented historically for the state picture not to be a drag, so it will be interesting to see if that holds up.  Is the double majority a sleeping curse that will wake up in the months to come or in the final results, or is it really going to be a non-issue this time around?  It turns out that if it is a non-issue, there's a reason for it, and that reason is Queensland.

With Newspoll on a brief hiatus following recent staff upheavals the recent national offerings have come from Essential, Resolve and new national player Redbridge.  Essential (2-5 August) continued to record slightly better results for Yes than other pollsters (excluding those I've dismissed for skewed wording), but still had Yes behind 43-47 with 10% undecided (a reversal of its previous 47-43-10 result).  Resolve (9-13 August) had Yes on 46-54 by forced choice (down from 48-52) and trailing 37-45-18 with undecided included, compared to 36-42-22 the month before.  (Resolve's undecided rate is higher because it does not include a second non-forcing prompt of the sort Essential and Newspoll include).  Redbridge (21-27 July) presented just a forced choice result which was 44-56, the worst so far from a pollster whose methods I considered to be acceptable.  After hearing a list of reasons for or against (the No list being slightly longer and a fair bit punchier) the result became 41-59.  

There was some confusion about the nature of the Redbridge poll with the Daily Telegraph describing it as "conducted exclusively for News Corp" while the pollster states that they paid for it themselves.  Both these things could in a sense be true if there was an arrangement to provide an exclusive to a particular outlet (presumably in return for publicity).  Redbridge would be required to disclose if there was a sponsor so it seems that there wasn't.

There are two main issues with aggregating the current state of play.  The first is that support for Yes has been falling so fast that assuming a continuing accelerating loss of support means Yes will have dropped more than another point since Resolve went out of the field and might even be below any poll so far released already.  But I'm reluctant to assume that especially when Simon Jackman's fancier tracking suggests the decline may have now stopped accelerating.  The other is what to do with Redbridge, which unlike other rare entrants so far is a bit off trend and hence may have a house effect.  I could assume one and in effect model it out of existence but I prefer to wait for a second data point which I believe is coming in a few weeks.  For the time being, then, I'm putting Yes on 45.0, but that's as of the most recent data nine days ago.

Key to colours: Green (previously red) - Newspoll, Magenta - Resolve, Yellow - Essential, Dark blue - JWS, Light blue - Freshwater, Black - Morgan, Red - Redbridge.  

Other noises

There were polling-shaped noises surrounding polls "commissioned by senior Liberal members" (which sounds like not real party internals as such, more likely people buying polling off their own bat) that were said to show a 34-47 result in Wentworth and a 38-56 result in Warringah.  The first was by an unstated pollster with unstated methods and questions and in general an electorate sample size of 500 is pretty useless unless it uses advanced respondent targeting.   

The second was slightly larger and by robopollster KJC Research as "part of a broader questionnaire" - this pollster under other names is often associated with rather long internal Liberal polls.  What is known is that the question here asked about “altering the Australian Constitution to specifically establish an Aboriginal and Torres Strait Islander Voice?” (my emphasis).  This was asked alongside a question about recognition, which is presented first in the report - if the recognition question was asked first then the ordering is likely to have caused respondents to say that they prefer recognition over a Voice.  The use of the word "specifically" could also create an impression that inserting the Voice is not recognition, but it clearly is recognition and not just a Voice in the context of the referendum question and the proposed text to be added to the Constitution.  The use of the word "specifically" alone is therefore sufficient reason to dismiss this polling, as at best it is measuring respondents' views on something that isn't happening,

Just today both the SMH and Daily Telegraph, within minutes of each other reported the Yes camp as believing that WA alongside Queensland was toast, that Yes was going well in NSW, Victoria and Tasmania and that SA was seen as crucial.  This was accompanied in the former case of claims from a Labor source that the Yes camp had internal polling with Yes at 48 and 49 nationally.  Such claims may be being circulated simply for the purposes of maintaining morale given the widespread coverage of the Voice as sliding towards heavy defeat.  It's also possible these are completely genuine internal polling results but even if so the difference between them and the public polls could be a matter of wording or other methods - as usual, such vague "internal polling" claims are best noted for the record only.

State picture 

I have carried on with the same methods as in previous editions (with minor adjustment for Redbridge issuing a partial breakout) and I now have the following aggregated estimates.  Figure in brackets is divergence from the national estimate:

VIC 49.0 (+4)

TAS 47.3 (+2.3)

NSW 45.5 (+0.5)

SA 45.1 (+0.1)

WA 43.2 (-1.8)

Qld 39.2 (-5.8)

Thus while Yes is currently behind in this polling-based estimate in every state, if Yes can gain 5% across the board (in the form of real gains, polling error or both) then Yes wins.  But history says that's an unusual script because ...

The Double Majority Handicap Stakes

This section is rated Wonk Factor 3/5; beware, number-crunching ahead!

I thought I'd have a look at how much of a hurdle the double majority has historically been.  Out of 44 referendum questions to date, eight have won both the national vote and four or more states, five have won the national vote but not enough states, none have won four or more states but lost the national vote, and the other 31 have failed on both criteria.  The record of five referendums losing on the state rule while winning the national vote suggests that winning four states is harder usually than winning 50% nationwide.

As one way of looking at this, I looked at the history of how the vote in the fourth most supportive state in referendums relates to the national vote.  Here's a graph of this, with successful referendums in green and failed referendums in red:

The regression line shows that the fourth-state vote has tended to run a bit below the national vote.  To put numbers on this, the median difference is 3.1%, meaning that a uniform swing to get Yes to 53.1% would be necessary for the fourth state to reach 50%.  The mean difference is 3.9%.  If I limit it to the more competitive referendums (which I've defined as those with Yes between 40 and 65 nationally, ignoring the lopsided cases) then the median is 2.8% and the mean is 3.4%.  Overall, historically, on average something like a 53% national vote is therefore needed for a 50-50 chance of a referendum passing.  

What is especially striking here is that only twice since federation has the fourth state beaten the national vote.  The only case of it doing so substantially is the green dot way above the regression line above.  That's the 1910 State Debts referendum, rejected 2 to 1 by New South Wales and passed overwhelmingly everywhere else.  The other case where the fourth state beat the national vote was by 0.02%, which was Surplus Revenue 1910 on the same day, and it didn't help there because both the national vote and the state vote lost.

There were another ten cases where the fourth state trailed the national vote by less than 1%, but the last of these was in 1951.  Four including the 1967 Aboriginals referendum had a gap of 1-2%, six had 2-3%, six had 3-4% and so on.  Four had gaps of more than 10%, including "Simultaneous Elections" 1977 which thereby failed with a Yes vote of 62.2% (it's the red outlier below).

Partly this is because if you pick six completely random numbers and take their average, there's a rather high chance that you won't get four above the average.  You'll get a lot of 3-3 splits, and when there are four or more on one side it could be the below side rather than the above side.  But more than that, referendums are a weighted average and historically the big states tend to vote Yes more than the smaller states.  This has become pronounced in more recent decades, such that in the 18 referendums in the last 50 years, NSW and Victoria have been in the top three Yes votes every time, and Tasmania has been in the bottom three every time with the lowest Yes vote in twelve of those.  It is, however, worth noting that the 1967 Aboriginals referendum was the most recent time Tasmania was not in the bottom three Yes states.  

The question then is whether the polling that suggests the state breakdown is really not a problem by itself for the Yes campaign is correct, or whether some invisible hand of referendums past will rise up and alter the state numbers.  I suggest a reason why the current relative state breakdowns might be correct is that something very similar happened in the 2022 federal election.

Federal election 2PPs, as opposed to referendums, show something more like the sort of pattern that would be expected randomly in terms of how the fourth best state compares to the national total.  On average (mean) the Labor 2PP in Labor's fourth-best state has run 1.3% behind Labor's national 2PP.  On average the Coalition's 2PP in the Coalition's fourth-best state has run 0.8% behind the Coalition's national 2PP, so there is not much skew either way.

But in 2022, Labor's fourth-best state, South Australia, beat the national Labor 2PP by 1.84%, which is the largest or very nearly the largest margin (there are precision issues with earlier elections) by which this has ever happened on either side.   And the gap between the Coalition's best state (Queensland) and its next best, at 5.47%, was the third largest for the Coalition ever, behind 1929 (Queensland, 6.4%) and 1983 (Tasmania, 5.88%, Franklin Dam.)

The Voice polling is echoing the 2022 federal picture of Queensland as now a very different place politically to the rest of the country.  If that carries through to the Voice then it's quite believable that the Yes vote can keep running ahead of its national average in four states.  That will be no good, however, if that national average stays so low or goes even lower. 

(NB In the marriage law postal survey, the fourth state beat the national average by 0.9% because the weakest state for Yes (NSW) was the largest.)